Amazon's keyword search is quite broken, and vendors exploit that. I have an HN comment somewhere where I was talking about an experiment I did; I took some of

Amazon's keyword search is quite broken, and vendors exploit that. I have an HN comment somewhere where I was talking about an experiment I did; I took some of the descriptions that exploit keyword matching, and asked GPT-3 my actual question. I think the specific example was waterproof LED strips. I noticed that many LED strips had the description "not waterproof" somewhere in the product name. If you search Amazon for "waterproof LED strip" (which BTW, I don't think is an actual thing), it returns all of the self-declared "non-waterproof" items, which is the exploit. I set up a GPT-3 completion where I asked questions like "is an Amazon item with the description 'foobar 12V LED light strip non-waterproof' waterproof?" For everthing I tested, GPT-3 returned completions like "No, of course it's not waterproof." If Amazon used ML like this to return search results, it would be slightly harder for vendors to exploit, and search would be much better. (But of course, sometimes it would return nothing when the user would have been happy with a non-waterproof LED light strip. I think that's a good thing, their Revenue Department probably disagrees though. If someone who wants waterproof lights gets one that's not waterproof and doesn't return it after paying their credit card bill, then "the system works".)